Transfer learning for improved LIBS analytical performance

abstract

English

Owing to its non-linear signal response, quantitative analysis via laser-induced breakdown spectroscopy (LIBS) is generally challenging. A major source of this non-linearity is the compounding impact of various matrix effects. Consequently, reliable quantitative LIBS analysis is generally performed by constructing a regression model using spectra from matrix-matched calibration standards of known compositions. Moreover, due to the high sensitivity of LIBS to the instrumental and experimental parameters, the calibration data generally must be collected under identical conditions as the target application. Hence, even relatively small changes in the LIBS apparatus often prompts the repeated collection of the calibration dataset. A notable example are the currently active Mars Rovers’ LIBS instruments. Namely, an extensive dataset comprising spectra of 406 calibration targets was constructed for the older ChemCam’s LIBS instrument. Recently, the SuperCam’s LIBS instrument’s calibration required the collection of a new calibration dataset, which consists of spectra of 334 calibration targets. Meanwhile, the regression models used for the SuperCam instrument could not benefit from the available data collected by the ChemCam instrument. In this work, we took advantage of the partially overlapping calibration datasets. Namely, we trained an artificial neural network to transform the ChemCam dataset into a form compatible with the SuperCam instrument. Consequently, the data available for training the regression model for the SuperCam instrument was considerably extended. This lead to the improvement of the calibration model’s performance in terms of root mean squared error of prediction of the major oxide content of the testing dataset.

machine learning; transfer learning; spectroscopic data; convolutional neural networks; Mars rovers

KÉPEŠ, E.; VRÁBEL, J.; POŘÍZKA, P.; KAISER, J.

2. 10. 2022